On 15/11/2017 16:11, Måns Rullgård wrote:
> Well, that's not good. I'll see if I can replicate it later this week.
In my latest test, I started from v4.14 and only applied the
following patch:
diff --git a/drivers/net/ethernet/aurora/nb8800.c
b/drivers/net/ethernet/aurora/nb8800.c
index e94159507847..5c109cc4bde8 100644
--- a/drivers/net/ethernet/aurora/nb8800.c
+++ b/drivers/net/ethernet/aurora/nb8800.c
@@ -317,6 +317,7 @@ static int nb8800_poll(struct napi_struct *napi, int budget)
len = RX_BYTES_TRANSFERRED(rxd->report);
+ udelay(200);
if (IS_RX_ERROR(rxd->report))
nb8800_rx_error(dev, rxd->report);
else
@@ -1246,10 +1247,12 @@ static int nb8800_hw_init(struct net_device *dev)
nb8800_writeb(priv, NB8800_PQ1, val >> 8);
nb8800_writeb(priv, NB8800_PQ2, val & 0xff);
+#if 0
/* Auto-negotiate by default */
priv->pause_aneg = true;
priv->pause_rx = true;
priv->pause_tx = true;
+#endif
nb8800_mc_init(dev, 0);
diff --git a/drivers/net/ethernet/aurora/nb8800.h
b/drivers/net/ethernet/aurora/nb8800.h
index aacc3cce2cc0..a2d4b841306a 100644
--- a/drivers/net/ethernet/aurora/nb8800.h
+++ b/drivers/net/ethernet/aurora/nb8800.h
@@ -8,7 +8,7 @@
#include <linux/clk.h>
#include <linux/bitops.h>
-#define RX_DESC_COUNT 256
+#define RX_DESC_COUNT 16
#define TX_DESC_COUNT 256
#define NB8800_DESC_LOW 4
Once I booted the SMP8759 board, I ran
ip addr add 172.27.64.23/18 brd 172.27.127.255 dev eth0
ip link set eth0 up
sleep 5
iperf3 -s
And from the PC, I ran
iperf3 -c 172.27.64.23 -u -b 0 -l 1400 -t 120
[ 5] local 172.27.64.23 port 5201 connected to 172.27.64.1 port 53771
[ ID] Interval Transfer Bandwidth Jitter Lost/Total
Datagrams
[ 5] 0.00-1.00 sec 5.89 MBytes 49.4 Mbits/sec 0.044 ms 77062/81472
(95%)
[ 5] 1.00-2.00 sec 6.25 MBytes 52.5 Mbits/sec 0.052 ms 80112/84796
(94%)
[ 5] 2.00-3.00 sec 6.26 MBytes 52.5 Mbits/sec 0.054 ms 80112/84797
(94%)
[ 5] 3.00-4.00 sec 6.26 MBytes 52.5 Mbits/sec 0.053 ms 80128/84814
(94%)
[ 5] 4.00-5.00 sec 6.26 MBytes 52.5 Mbits/sec 0.049 ms 80112/84798
(94%)
[ 5] 5.00-6.00 sec 6.26 MBytes 52.5 Mbits/sec 0.052 ms 80096/84783
(94%)
[ 5] 6.00-7.00 sec 6.26 MBytes 52.5 Mbits/sec 0.047 ms 80096/84784
(94%)
[ 5] 7.00-8.00 sec 6.26 MBytes 52.5 Mbits/sec 0.060 ms 80128/84815
(94%)
iperf3: OUT OF ORDER - incoming packet = 731264 and received packet = 0 AND SP
= 731279
iperf3: OUT OF ORDER - incoming packet = 731265 and received packet = 0 AND SP
= 731279
iperf3: OUT OF ORDER - incoming packet = 731266 and received packet = 0 AND SP
= 731279
iperf3: OUT OF ORDER - incoming packet = 731267 and received packet = 0 AND SP
= 731279
iperf3: OUT OF ORDER - incoming packet = 731268 and received packet = 0 AND SP
= 731279
iperf3: OUT OF ORDER - incoming packet = 731269 and received packet = 0 AND SP
= 731279
iperf3: OUT OF ORDER - incoming packet = 731270 and received packet = 0 AND SP
= 731279
iperf3: OUT OF ORDER - incoming packet = 731271 and received packet = 0 AND SP
= 731279
iperf3: OUT OF ORDER - incoming packet = 731272 and received packet = 0 AND SP
= 731279
[ 5] 8.00-9.00 sec 4.17 MBytes 35.0 Mbits/sec 0.149 ms 53104/56220
(94%)
[ 5] 9.00-10.00 sec 0.00 Bytes 0.00 bits/sec 0.149 ms 0/0 (0%)
[ 5] 10.00-11.00 sec 0.00 Bytes 0.00 bits/sec 0.149 ms 0/0 (0%)
[ 5] 11.00-12.00 sec 0.00 Bytes 0.00 bits/sec 0.149 ms 0/0 (0%)
[ 5] 12.00-13.00 sec 0.00 Bytes 0.00 bits/sec 0.149 ms 0/0 (0%)
[ 5] 13.00-14.00 sec 0.00 Bytes 0.00 bits/sec 0.149 ms 0/0 (0%)
^C
- - - - - - - - - - - - - - - - - - - - - - - - -
[ ID] Interval Transfer Bandwidth Jitter Lost/Total
Datagrams
[ 5] 0.00-17.04 sec 0.00 Bytes 0.00 bits/sec 0.149 ms 690950/731279
(94%)
[SUM] 0.0-17.0 sec 9 datagrams received out-of-order
iperf3: interrupt - the server has terminated
# ping -c 10 172.27.64.1
PING 172.27.64.1 (172.27.64.1): 56 data bytes
--- 172.27.64.1 ping statistics ---
10 packets transmitted, 0 packets received, 100% packet loss
Given the out-of-order datagrams, I'm wondering if it's possible
for the DMA engine to overwrite a not-yet-read descriptor?
The EOC flag should stop the DMA engine though...
Maybe some kind of race...
I don't think I've been able to trigger the wedge when 256 descriptors
are used.
